llm-kb
← Back to news
Incident

LLM-as-judge eval misses 3% of users due to unseen output format

A developer recounts how their CPO mandated LLM eval automation using GPT-4o as a judge with an 8-dimension rubric. After three months of success, a system prompt tweak caused the judge to miss a completely different output format for 3% of users, leading to undetected regressions discovered via support tickets.

11 engagement·1 source·Fri, Jul 10, 2026, 08:21 PM
The developer's CPO mandated LLM eval automation in November after a conference talk. The developer set up GPT-4o as a judge with an 8-dimension rubric, running on every deploy. For the first three months it worked, catching a couple obvious regressions. In December, the ML lead tweaked a system prompt to improve one specific edge case. The judge scored it 8.7/10, so they shipped. However, about 3% of users were in a flow that triggered a completely different output format the judge had never seen in training examples, so it scored fine. The issue was found from support tickets on Monday morning.

Entities

GPT-4o(model)CPO(person)ML lead(person)OpenAI(company)

Related

CommunitySat, Jul 11, 2026, 06:13 PM

Developer asks community for agent evaluation practices, cites silent breakage

A developer building AI agents reports that prompt or MCP changes often break silently despite passing manual tests. They ask the community about evaluation methods, including fixed test cases, skill-level vs. end-to-end checks, and tools like DeepEval, LangSmith, and Ragas.

10 engagement·1 source·reddit
CommunitySun, Jul 12, 2026, 07:33 AM

User discovers that describing desired output quality outperforms step-by-step instructions in prompts

A Reddit user reports that shifting from detailed step-by-step instructions to describing the desired outcome (e.g., 'a great version would make a busy person understand the tradeoff in ten seconds') dramatically improves LLM output quality. The post highlights that models are better at navigating to a well-defined finish line than following clumsy instructions.

3 engagement·1 source·reddit
IncidentSat, Jul 11, 2026, 04:43 PM

Developer discovers chatbot quality degrades after 5 turns

A developer reports that their chatbot, which passes quality evals on short interactions, gradually loses context after about 5 turns, forgetting user constraints and contradicting itself. This highlights a common limitation in current conversational AI systems.

3 engagement·1 source·reddit
PaperThu, Jul 9, 2026, 05:01 PM

arXiv paper benchmarks LLM judges for citation quality in deep-research systems

A new arXiv paper studies the calibration of LLM judges used as reward models in reinforcement learning for citation quality in deep-research systems. The work evaluates how capable and biased an LLM judge must be to reliably score rubric criteria like source relevance and factual support for attribution-citation pairs. This matters for practitioners building RL-based systems that depend on automated citation verification.

0 engagement·1 source·arxiv
arXiv
BenchmarkWed, Jul 8, 2026, 01:00 PM

OpenAI analysis reveals flaws in SWE-Bench Pro coding benchmark

OpenAI published an analysis uncovering reliability issues in SWE-Bench Pro, a popular benchmark for evaluating AI coding models. The findings raise concerns about the accuracy of benchmark scores, potentially affecting how developers and researchers trust model evaluations.

0 engagement·1 source·rss
RSS